Abstract: Rich knowledge can be acquired by mining the network topology, especially the various sub-structures that each node lies in. Different structures are formed by different potential interactions, which on the other hand indicates different effects between the centric node and its surroundings. In fact, the interaction between the neighbor structures and the mutual influence between the nodes are cause and effect. This paper aims at three most common neighbor structures and proposes DNSI (neighbor structure influence based on deep learning) to analyze the network neighbor structure influence combined with deep learning. By extracting features of the network data in the image format, DNSI can obtain the influence of three neighbor structures. This paper conducts this model on several real-world web data sets to predict node properties, measures category nodes centrality and predicts user behaviors. The experimental results show that this model has superiority in the vast majority of cases.

Key words: social networks, neighbor structure, node properties, centrality measurement, behavior prediction

摘要： 通过研究网络的拓扑结构可以探索到丰富的知识，特别是网络中节点的邻居可以形成不同的邻居结构，而不同的结构蕴含着不同的意义，进而也有着不同的影响。实际上，邻居结构与节点的交互行为之间是互相影响、互为因果的。对三种最为普遍的邻居结构进行分析，并提出结合深度学习的网络邻居结构影响力模型DNSI（neighbor structure influence based on deep learning）。通过对图片格式的网络数据提取特征，DNSI可以得到三种邻居结构影响力。分别在几个真实世界网络数据集上进行节点属性预测、类别中心度度量和用户行为预测等任务，实验结果表明该模型在绝大多数情况下具有优越性。

关键词: 社交网络, 邻居结构, 节点属性, 中心度度量, 行为预测